Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

ChatGPT outperforms crowd-workers for text-annotation tasks

arxiv.org

Post reply on HN
Print view View original

121–130 of 206 posts

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#121

Post by cormacrelf »

Imnimo wrote: post #20

My main take away here is that Turkers are terrible at some of these tasks. The "stance" task is, "Classify the tweet as having a positive stance towards Section 230, a negative stance, or a neutral stance.", and the Turkers accuracy was like 20%. Even in its best task, ChatGPT only got 75% accuracy.

That’s in the nature of the question, which is probably not representative of what people might want a sentiment analysis for. Section 230 has been the subject of what can only be described as a massive disinformation campaign. Most of the participants were not aware. But ask anybody today what the law even means and you will get 99% completely wrong answers. So even neutral factual statements will be misinterpreted if people think Section 230 is some nebulous but definitely bad thing and so anything about it is probably bad.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#122

Post by YeGoblynQueenne »

Suppose you have two classifiers, A and B, and some un-annotated data, D. You want to know how good is classifier B at annotating the data, compared to classifier A.

One problem is that you don't have the ground truth for D. So you start by annotating D with the labels assinged by a third classifier, C:

  C(D) → D₁
Having thus established a modicum of "ground truth", ish, you proceed to annotate D with the two classifiers you are comparing, A and B:

  A(D) → D₂
  B(D) → D₃
Then you compare the classifications D₂ and D₃ to D₁, and find that D₃ better approximates D₁ than D₂.

What is the result of the experiment? I summarise it as follows:

  Classifier B better approximates the labelling of D by classifier C than
  classifier A.
Now we can name the three classifiers as they were used in the experiments in the linked article:

A: Human annotators.

B: ChatGPT

C: Human annotators.

So the result of the paper is that, plugging in the names:

  ChatGPT better approximates the labelling of D by Human annotators than
  Human annotators.
And that, is the finding of the paper.

Which is clearly absurd and a cause to re-think methodology.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#123

Post by dotancohen »

CGamesPlay wrote: post #110
renewiltord wrote: post #108

Earlier quoted context omitted.

You always use multiple MTurks on same job when you use them.

OK, but if all MTurks are doing this (because why wouldn't they) then you are just sampling a random variable.

And you'll thus see that the answers converge at the rate a random variable would. Whereas with more accurate methods, the answers will generally agree with one another.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#124

Post by YeGoblynQueenne »

nico wrote: post #38
jacknews wrote: post #16

So is this 'AI trains AI better than people can'? And presumably the better-trained AI will also be better again at training. I think I've seen this movie.

That's exactly what's coming. AI will train AI. At some point AI is going to stop needing us to keep evolving. Not sure what that will be like.

Since I've been seeing this wild fantasy being bandied about for a while now, I have to point out that, if you had "AI" that could train "AI", you wouldn't need to train any more "AI". Because at that point, there would be nothing to gain.

Suppose you have a text classifier that can produce text classifications just as good as those of human annotators, so that you could use it to train other text classifiers. At that point, why would you need to train other text classifiers? What changes to predictive accuracy would you expect to achieve?

Or suppose you had a language model that could produce language just as well as humans (not just in terms of grammaticality, but also in terms of making sense, I guess?). At that point, why would you need to use the text generated by that language model to train another language model? What kind of changes in the quality of the new model's generated text would you expect to see?

Same goes for image classifiers, and any other classifiers, or generative models you may care to consider.

Note that both language models and image classifiers have been "beating" human performance in benchmarks for a while now, and still they are not used to train other classifiers. And the reason for that is that it doesn't make sense: the ground truth is always the decisions made by humans. There is nothing in machine learning theory or practice that says that a classifier can perform better than whatever process originally labelled its data. If it looks like it does, it's because there's something wrong with the methodology (i.e. overfitting, bad benchmarks, some other kind of nonsense).

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#125

Post by dotancohen »

raincole wrote: post #115
quickthrower2 wrote: post #104

Earlier quoted context omitted.

Probability: always counterintuitive

It's just how you frame the question (in this parcitular case). If random guess were not 33% correct, you effectively found a way to "win" in rock-paper-scissor.

I do have a strategy for always winning rock-paper-scissors.

Play often for unimportant reasons, and always choose "rock". Now that your adversaries are conditioned to expect you to choose "rock", when an important decision is being decided choose "scissors".

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#126

Post by sebzim4500 »

YeGoblynQueenne wrote: post #122

Suppose you have two classifiers, A and B, and some un-annotated data, D. You want to know how good is classifier B at annotating the data, compared to classifier A. One problem is that you don't have the ground truth for D. So you start by annotating D with the labels assinged by a third classifier, C: C(D) → D₁ Having thus established a modicum of "ground truth", ish, you proceed to annotate D with the two classifi…

I think your assumption that all humans are equally capable should be reconsidered.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#127

Post by ninjin »

swyx wrote: post #11
ninjin wrote: post #5

Earlier quoted context omitted.

Obviously not. If we had a general solution to language intelligence we would have artificial intelligence at the level of at least human intelligence – which we do not . Rather, the right question to ask is which language intelligence tasks currently have acceptable performance and under which conditions (text domain, etc.). Clearly this is a much more difficult question and with a lot more nuance to it, even if it…

as a senior academic in the area, might you have a list of the most influential papers in your field in the past year that you would recommend?

I am actually not quite sure what would be best to recommend, as the pandemic has seen me lag behind the zeitgeist somewhat. A recent favourite would be the Toolformer paper [1], that I intend to read in detail later today. If an LLM would be able to use external tools efficiently, it could be rather powerful and perhaps allow us to scale down the parameter sizes, which somewhat fascinates me.

[1]: https://arxiv.org/abs/2302.04761

Other research questions, but without concrete papers to reference to that current keeps me up at night: 1.) to which degree can we train substantially smaller LLMs for specific tasks that could be run in-house, 2.) it seems like these new breakthroughs may need a different mode of evaluation compared to what we have used since the 80s in the field and I am not sure what that would look like (maybe along the lines of HELM [2]?), and 3.) can AI academia continue to operate in the way it currently does with small research teams or is a change towards what we see in the physical sciences necessary?

[2]: https://arxiv.org/abs/2211.09110

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#128

Post by YeGoblynQueenne »

ninjin wrote: post #5
sorz wrote: post #3

Is NLP a solved problem now?

Obviously not. If we had a general solution to language intelligence we would have artificial intelligence at the level of at least human intelligence – which we do not . Rather, the right question to ask is which language intelligence tasks currently have acceptable performance and under which conditions (text domain, etc.). Clearly this is a much more difficult question and with a lot more nuance to it, even if it…

This is clearly preliminary work. Not to be disparaging of the authors' background but their background is in political sciences, not in machine learning or NLP which should account for the limitations of the study. But anyway that's just an arxiv preprint so probably more like something exploratory than a research direction the authors are invested in.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#129

Post by YeGoblynQueenne »

famouswaffles wrote: post #72
shock-value wrote: post #68

Earlier quoted context omitted.

> ... these large language models are already intelligent enough to matter. I'm definitely not contesting that. I've always considered the idea of "AGI" to mean something of the holy grail of machine learning -- the point at which there is no real point in pursuing further advances in artificial intelligence because the AI itself will discover and apply such augmentations using its own capabilities. I have seen no ev…

See what you're describing is much closer to ASI. At least, it used to be. This is the big problem I have. The constant post shifting is maddening. AGI went from meaning Generally Intelligent to as smart as Human experts and then now smarter than all experts combined. You'll forgive me if I no longer want to play this game. I know some researchers disagree. That's fine. The point I was really getting at is that no re…

>> The point I was really getting at is that no researcher worth his salt can call these models narrow anymore.

Are you talking about large language models (LLMs)? Because those are narrow, and brittle, and dumb as bricks, and I don't care a jot about your "No True Scotsman". LLMs can only operate on text, they can only output text that demonstrates "reasoning" when their training text has instances of text detailing the solutions of reasoning problems similar to the ones they're asked to solve, and their output depends entirely on their input: you change the prompt and the "AGI" becomes a drooling idiot, and v.v.

That's no sign of intelligence and you should re-evaluate your unbridled enthusiasm. You believe in magick, and you are loudly proclaiming your belief in magick. Examples abound in history that magick doesn't work, and only science does.

Re: ChatGPT outperforms crowd-workers for text-annotation tasks

#130

Post by YeGoblynQueenne »

sebzim4500 wrote: post #126
YeGoblynQueenne wrote: post #122

Suppose you have two classifiers, A and B, and some un-annotated data, D. You want to know how good is classifier B at annotating the data, compared to classifier A. One problem is that you don't have the ground truth for D. So you start by annotating D with the labels assinged by a third classifier, C: C(D) → D₁ Having thus established a modicum of "ground truth", ish, you proceed to annotate D with the two classifi…

I think your assumption that all humans are equally capable should be reconsidered.

That seems to be the authors' assumption.
Post reply on HN